Two things about new engineers seem like separate observations. The first is that their confusion is the only honest reading you will ever get of how legible your system is, and that the reading expires in about three weeks. The second is that once the confusion is behind them, the same person quietly stops being an advocate for fixing it — the cost is sunk for them now, and the fix is not. They look like two different problems, one about measurement and one about incentives.

They’re the same fact seen twice. Comprehension is the only significant cost in a software system that is never metered.

Everything else gets a number. Build times go on a dashboard. Query latency has a percentile and an alert. Cloud spend arrives as an invoice that somebody has to explain at the end of the month. Even the vague costs get proxies — velocity for delivery, incident count for reliability. But the hours a person spends assembling a private working model of why there are two deploy steps produce no artifact at all. No ticket, no metric, no line item. The work happens, the money is spent, and nothing anywhere records that it happened.

An unmetered cost has two properties, and each one produces one of the observations. Because it’s never aggregated, it’s only visible at the moment of payment — which is exactly why the new person is an instrument and everyone else isn’t: they’re the only one currently being charged. And because it’s charged per person rather than once, it never accumulates into a number large enough to argue with. Twenty people paying thirty hours each is six hundred hours, and it looks like nothing, because it was never a line anywhere. It was twenty separate weeks that each felt normal.

The comparison that makes this concrete: a system that burned an extra six hundred hours of compute a year would be someone’s actual assignment by Thursday. A system that burns six hundred hours of human comprehension a year is described, when it’s described at all, as having a bit of a learning curve. The second cost is more expensive per hour than the first and recurs with every hire, and it is the one that reliably has no owner.

So the practical move isn’t a cultural exhortation to be more welcoming to new people. It’s bookkeeping: make the cost leave a trace. Write down every question asked in someone’s first month — the question, not the answer — and let the repeats accumulate into a document that isn’t onboarding material but a defect list. Note when a fix would have made a question unnecessary. That’s it. You’re not creating a new cost; you’re producing the first record of one you’ve been paying all along, in a form somebody can eventually point at.

Words on a wiki won’t change the incentives on their own. But the reason the confusing thing survives isn’t that anyone decided it was worth the price. It’s that nobody has ever seen the price. Being unmeasured is the whole of its protection.